Spaces:
Sleeping
Sleeping
File size: 11,365 Bytes
3fef185 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from src.text_embedding import TextEmbeddingModel
class ClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, in_dim, out_dim):
super(ClassificationHead, self).__init__()
self.dense1 = nn.Linear(in_dim, in_dim//4)
self.dense2 = nn.Linear(in_dim//4, in_dim//16)
self.out_proj = nn.Linear(in_dim//16, out_dim)
nn.init.xavier_uniform_(self.dense1.weight)
nn.init.xavier_uniform_(self.dense2.weight)
nn.init.xavier_uniform_(self.out_proj.weight)
nn.init.normal_(self.dense1.bias, std=1e-6)
nn.init.normal_(self.dense2.bias, std=1e-6)
nn.init.normal_(self.out_proj.bias, std=1e-6)
def forward(self, features):
x = features
x = self.dense1(x)
x = torch.tanh(x)
x = self.dense2(x)
x = torch.tanh(x)
x = self.out_proj(x)
return x
class SimCLR_Classifier_SCL(nn.Module):
def __init__(self, opt,fabric):
super(SimCLR_Classifier_SCL, self).__init__()
self.temperature = opt.temperature
self.opt=opt
self.fabric = fabric
self.model = TextEmbeddingModel(opt.model_name)
self.device=self.model.model.device
if opt.resum:
state_dict = torch.load(opt.pth_path, map_location=self.device)
self.model.load_state_dict(state_dict)
self.esp=torch.tensor(1e-6,device=self.device)
self.classifier = ClassificationHead(opt.projection_size, opt.classifier_dim)
self.a=torch.tensor(opt.a,device=self.device)
self.d=torch.tensor(opt.d,device=self.device)
self.only_classifier=opt.only_classifier
def get_encoder(self):
return self.model
def _compute_logits(self, q,q_index1, q_index2,q_label,k,k_index1,k_index2,k_label):
def cosine_similarity_matrix(q, k):
q_norm = F.normalize(q,dim=-1)
k_norm = F.normalize(k,dim=-1)
cosine_similarity = q_norm@k_norm.T
return cosine_similarity
logits=cosine_similarity_matrix(q,k)/self.temperature
q_labels=q_label.view(-1, 1)# N,1
k_labels=k_label.view(1, -1)# 1,N+K
same_label=(q_labels==k_labels)# N,N+K
#model:model set
pos_logits_model = torch.sum(logits*same_label,dim=1)/torch.max(torch.sum(same_label,dim=1),self.esp)
neg_logits_model=logits*torch.logical_not(same_label)
logits_model=torch.cat((pos_logits_model.unsqueeze(1), neg_logits_model), dim=1)
return logits_model
def forward(self, batch, indices1, indices2,label):
bsz = batch['input_ids'].size(0)
q = self.model(batch)
k = q.clone().detach()
k = self.fabric.all_gather(k).view(-1, k.size(1))
k_label = self.fabric.all_gather(label).view(-1)
k_index1 = self.fabric.all_gather(indices1).view(-1)
k_index2 = self.fabric.all_gather(indices2).view(-1)
#q:N
#k:4N
logits_label = self._compute_logits(q,indices1, indices2,label,k,k_index1,k_index2,k_label)
out = self.classifier(q)
if self.opt.AA:
loss_classfiy = F.cross_entropy(out, indices1)
else:
loss_classfiy = F.cross_entropy(out, label)
gt = torch.zeros(bsz, dtype=torch.long,device=logits_label.device)
if self.only_classifier:
loss_label = torch.tensor(0,device=self.device)
else:
loss_label = F.cross_entropy(logits_label, gt)
loss = self.a*loss_label+self.d*loss_classfiy
if self.training:
return loss,loss_label,loss_classfiy,k,k_label
else:
out = self.fabric.all_gather(out).view(-1, out.size(1))
return loss,out,k,k_label
class SimCLR_Classifier_test(nn.Module):
def __init__(self, opt,fabric):
super(SimCLR_Classifier_test, self).__init__()
self.fabric = fabric
self.model = TextEmbeddingModel(opt.model_name)
self.classifier = ClassificationHead(opt.projection_size, opt.classifier_dim)
self.device=self.model.model.device
def forward(self, batch):
q = self.model(batch)
out = self.classifier(q)
return out
class SimCLR_Classifier(nn.Module):
def __init__(self, opt,fabric):
super(SimCLR_Classifier, self).__init__()
self.temperature = opt.temperature
self.opt=opt
self.fabric = fabric
self.model = TextEmbeddingModel(opt.model_name)
if opt.resum:
state_dict = torch.load(opt.pth_path,
map_location=self.model.device)
self.model.load_state_dict(state_dict)
self.device = self.model.model.device
self.esp = torch.tensor(1e-6,device=self.device)
self.a = torch.tensor(opt.a,
device=self.device)
self.b = torch.tensor(opt.b,
device=self.device)
self.c = torch.tensor(opt.c,
device=self.device)
self.classifier = ClassificationHead(opt.projection_size,
opt.classifier_dim)
self.only_classifier = opt.only_classifier
def get_encoder(self):
return self.model
def _compute_logits(self,
q,q_index1, q_index2, q_label,
k,k_index1,k_index2,k_label):
def cosine_similarity_matrix(q, k):
q_norm = F.normalize(q,dim=-1)
k_norm = F.normalize(k,dim=-1)
cosine_similarity = q_norm@k_norm.T
return cosine_similarity
logits=cosine_similarity_matrix(q,k)/self.temperature
q_index1=q_index1.view(-1, 1)# change to tensor of size N, 1
q_index2=q_index2.view(-1, 1)# change to tensor of size N, 1
q_labels=q_label.view(-1, 1)# change to tensor of size N, 1
k_index1=k_index1.view(1, -1)# 1,N+K
k_index2=k_index2.view(1, -1) #1, N+K
k_labels=k_label.view(1, -1)# 1,N+K
same_mixed = (q_index1== k_index1)
same_set=(q_index2==k_index2)# N,N+K
same_label=(q_labels==k_labels)# N,N+K
is_human=(q_label==1).view(-1)
is_machine=(q_label==0).view(-1)
is_mixed=(q_index1==1).view(-1)
#human: human
pos_logits_human = torch.sum(logits*same_label,dim=1)/torch.max(torch.sum(same_label,dim=1),self.esp)
neg_logits_human=logits*torch.logical_not(same_label)
logits_human=torch.cat((pos_logits_human.unsqueeze(1), neg_logits_human), dim=1)
logits_human=logits_human[is_human]
#human+ai: general
pos_logits_mixed = torch.sum(logits*same_mixed,dim=1)/torch.maximum(torch.sum(same_mixed,dim=1),self.esp)
neg_logits_mixed=logits*torch.logical_not(same_mixed)
logits_mixed=torch.cat((pos_logits_mixed.unsqueeze(1), neg_logits_mixed), dim=1)
logits_mixed=logits_mixed[is_mixed]
#human+ai: model
pos_logits_mixed_set = torch.sum(logits*torch.logical_and(same_mixed, same_set),dim=1)/torch.max(torch.sum(torch.logical_and(same_mixed, same_set),dim=1),self.esp)
neg_logits_mixed_set=logits*torch.logical_not(torch.logical_and(same_mixed, same_set))
logits_mixed_set=torch.cat((pos_logits_mixed_set.unsqueeze(1), neg_logits_mixed_set), dim=1)
logits_mixed_set=logits_mixed_set[is_mixed]
#model set:label
pos_logits_set = torch.sum(logits*same_set,dim=1)/torch.max(torch.sum(same_set,dim=1),self.esp)
neg_logits_set=logits*torch.logical_not(same_set)
logits_set=torch.cat((pos_logits_set.unsqueeze(1), neg_logits_set), dim=1)
logits_set=logits_set[is_machine]
#label: label
pos_logits_label = torch.sum(logits*same_label, dim=1)/torch.max(torch.sum(same_label,dim=1),self.esp)
neg_logits_label=logits*torch.logical_not(same_label)
logits_label=torch.cat((pos_logits_label.unsqueeze(1), neg_logits_label), dim=1)
logits_label=logits_label[is_machine]
return logits_human, logits_mixed, logits_mixed_set, logits_set, logits_label
def forward(self, encoded_batch, label, indices1, indices2):#, weights):
# print(len(text))
q = self.model(encoded_batch)
k = q.clone().detach()
k = self.fabric.all_gather(k).view(-1, k.size(1))
k_label = self.fabric.all_gather(label).view(-1)
k_index1 = self.fabric.all_gather(indices1).view(-1)
k_index2 = self.fabric.all_gather(indices2).view(-1)
#q:N
#k:4N
logits_human, logits_mixed, logits_mixed_set, logits_set, logits_label = self._compute_logits(q,indices1, indices2,label,
k,k_index1,k_index2,k_label)
out = self.classifier(q)
if self.opt.AA:
loss_classfiy = F.cross_entropy(out, indices1)
else:
loss_classfiy = F.cross_entropy(out, label) #, weight=weights)
gt_mixed = torch.zeros(logits_mixed.size(0),
dtype=torch.long,
device=logits_mixed.device)
gt_mixed_set = torch.zeros(logits_mixed_set.size(0),
dtype=torch.long,
device=logits_mixed_set.device)
gt_set = torch.zeros(logits_set.size(0),
dtype=torch.long,
device=logits_set.device)
gt_label = torch.zeros(logits_label.size(0),
dtype=torch.long,
device=logits_label.device)
gt_human = torch.zeros(logits_human.size(0),
dtype=torch.long,
device=logits_human.device)
loss_mixed = F.cross_entropy(logits_mixed,
gt_mixed)
loss_mixed_set = F.cross_entropy(logits_mixed_set,
gt_mixed_set)
loss_set = F.cross_entropy(logits_set,
gt_set)
loss_label = F.cross_entropy(logits_label,
gt_label)
if logits_human.numel()!=0:
loss_human = F.cross_entropy(logits_human.to(torch.float64),
gt_human)
else:
loss_human=torch.tensor(0,device=self.device)
loss = self.a*loss_set + (4*self.b-self.a)*loss_label + self.b*loss_human+ self.b*loss_mixed + \
2*self.b*loss_mixed_set+self.c*loss_classfiy
if self.training:
if self.opt.AA:
return loss,loss_mixed, loss_mixed_set,loss_set,loss_label,loss_human,loss_classfiy,k,k_index1
else:
return loss,loss_mixed, loss_mixed_set,loss_set,loss_label,loss_classfiy,loss_human,k,k_label
else:
out = self.fabric.all_gather(out).view(-1, out.size(1))
if self.opt.AA:
return loss,out,k,k_index1
else:
return loss,out,k,k_label
|